Monosidum glutamate umami intensity evaluation method and system based on EEG and machine learning

Abstract

The invention discloses a monosodium glutamate umami intensity evaluation method and system based on EEG and machine learning. The method comprises the following steps: acquiring brain wave data after testers taste monosodium glutamate solution samples with specific concentrations; preprocessing the brain wave data, and extracting brain wave features including a plurality of time domain features and frequency domain features; performing feature selection on the plurality of time domain features and frequency domain features of the brain waves by using a feature selection method based on mutual information to obtain a feature data set; dividing the feature data set into a training set and a test set; using the optimized random forest model to train the training set, and outputting the corresponding monosodium glutamate umami intensity; and predicting the umami intensity of the monosodium glutamate solution sample to be detected by using the trained random forest model. According to the method and the system, the umami intensity value of the monosodium glutamate can be relatively objectively and accurately obtained.

Description

Technical Field

[0001] The present invention relates to the fields of food science and artificial intelligence technology, and more particularly to a method and system for evaluating the umami flavor intensity of monosodium glutamate based on EEG and machine learning. Background Art

[0002] Monosodium Glutamate (MSG), whose main ingredient is monosodium glutamate (C5H8NNaO4), is a common food flavoring that enhances the umami flavor of food and adds depth to flavors. However, despite numerous scientific studies demonstrating the safety of moderate MSG consumption, some consumers remain concerned about its safety, citing the potential for excessive consumption to cause headaches, allergies, and other adverse reactions.

[0003] Therefore, establishing a scientific umami intensity evaluation system is crucial for the rational use of MSG and the long-term development of the food industry.

[0004] Currently, existing methods for evaluating umami intensity primarily rely on sensory analysis, lacking an objective and precise method for distinguishing the umami intensity of MSG solutions at varying concentrations. Therefore, developing a more accurate and objective umami intensity evaluation system is not only of great significance to the food industry but also contributes to the scientific use of MSG and other flavor enhancers.

[0005] Based on EEG technology and machine learning methods, this paper aims to develop an innovative umami intensity evaluation system to improve the scientificity and accuracy of umami analysis. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for evaluating the umami intensity of MSG based on EEG and machine learning, which are used to at least solve some of the technical problems in the background technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] The present invention first provides a method for evaluating the umami flavor intensity of MSG based on EEG and machine learning, comprising the following steps:

[0009] Obtain brain wave data after testers taste a sample of a specific concentration of MSG solution;

[0010] Preprocess the EEG data and extract EEG features, including multiple time domain features and frequency domain features;

[0011] A feature selection method based on mutual information is used to select multiple time domain features and frequency domain features of brain waves to obtain a feature data set;

[0012] Dividing the feature dataset into a training set and a test set;

[0013] The optimized random forest model is used to train the training set and output the corresponding MSG umami intensity;

[0014] The trained random forest model is used to predict the umami intensity of the MSG solution sample to be tested.

[0015] Furthermore, the brain wave data is preprocessed, specifically including the following steps:

[0016] Downsample the EEG data to 250Hz;

[0017] Filter the 250Hz EEG data, and set the filter range to 0.5-40Hz;

[0018] The filtered EEG data were time-stamped and segmented at 2-second intervals, with each segment of EEG data being treated as an independent sample.

[0019] Furthermore, EEG features are extracted, specifically including:

[0020] Extract the time domain features of the preprocessed EEG data, including the maximum signal amplitude, minimum signal amplitude, average signal amplitude, median signal amplitude, peak-to-peak value, signal variance, signal standard deviation, signal kurtosis, signal skewness, RMS amplitude, and average rectified signal value;

[0021] The preprocessed EEG data is converted into a frequency domain signal using Fourier transform, and multiple frequency domain features of the frequency domain signal are extracted, including frequency centroid, RMS frequency, frequency standard deviation, spectral skewness average, and power spectrum entropy.

[0022] Furthermore, a feature selection method based on mutual information is used to select multiple time domain features and frequency domain features of brain waves, which specifically includes the following steps:

[0023] Calculate the mutual information of all time domain features and frequency domain features of the same EEG data sample;

[0024] All the mutual information values obtained are sorted in descending order, and the first 30 features are selected as input features.

[0025] Furthermore, the mutual information of all time domain features and frequency domain features in the same EEG data sample is calculated, specifically including the following formula:

[0026]

[0027] Among them, P(x,y) represents the joint probability distribution of feature X and feature Y, and P(x) and P(y) are the marginal probability distributions of feature X and feature Y respectively.

[0028] Furthermore, during the training of the random forest model, GridSearchCV is used to optimize the model hyperparameters, specifically including the following formula:

[0029]

[0030] Where θ is a set of hyperparameters, including the number of features max_features and depth max_depth of the random forest model; K is k-fold cross validation (usually K = 5); X k ,Y k is the k-th fold training data and label; Score is the performance evaluation index of the random forest model (such as accuracy, F1 score, etc.), θ * is the optimal hyperparameter after optimization.

[0031] Furthermore, during the training process of the random forest model, a variance-based depth adjustment strategy is adopted to dynamically adjust the depth of the random forest model decision tree.

[0032] Furthermore, a variance-based depth adjustment strategy is used to dynamically adjust the depth of the decision tree of the random forest model, specifically including:

[0033] Calculate the variance of sample features in each decision tree node of the random forest model;

[0034] When the variance of the sample features in a decision tree node is less than the set stop threshold, the decision tree node stops splitting;

[0035] When the variance of the sample features in a decision tree node is greater than the set split threshold, the decision tree node continues to split.

[0036] The present invention also discloses a system for evaluating the umami intensity of MSG based on EEG and machine learning, comprising a computer system. When executed, the computer system can implement any one of the methods for evaluating the umami intensity of MSG based on EEG and machine learning described in the present invention.

[0037] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a method and system for evaluating the umami intensity of MSG based on EEG and machine learning, which has the following beneficial effects:

[0038] This method combines EEG and machine learning to segment and comprehensively extract features from raw EEG data, significantly improving model performance. Compared to traditional methods that directly use random forest models and train data based on power spectral density (PSD) across frequency bands, this method offers a more comprehensive approach to EEG feature extraction, incorporating not only traditional PSD features but also richer information, improving data representation.

[0039] In terms of algorithm optimization, the present invention adopts a feature selection method that deeply integrates mutual information (MI) and random forest. Unlike traditional MI feature selection, which only screens features based on global correlation, the present invention combines the decision tree splitting strategy of random forest in the feature selection process and dynamically adjusts the MI threshold so that the selected features not only have global correlation, but also can adapt to the local decision-making requirements of the model, thereby improving the generalization ability and computational efficiency of the model. The present invention further introduces a dynamic adjustment mechanism based on variance to adaptively optimize the depth of the decision tree during the growth of the decision tree, avoid overfitting of the model and improve stability.

[0040] This paper adopts an improved GridSearchCV hyperparameter optimization strategy. It not only automatically searches for the optimal hyperparameter combination during the cross-validation (CV) process, but also combines the feedback mechanism of the feature selection module to enable the hyperparameter search to dynamically adjust the MI screening threshold and decision tree depth parameters, achieving collaborative optimization of feature selection and model optimization. Experimental results show that this method can effectively improve classification performance, increasing the model prediction accuracy from 92.44% to 97.03%, significantly enhancing the classification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. The drawings described below are embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0042] Figure 1 The present invention provides a schematic diagram of the overall process of the method. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] The embodiment of the present invention discloses a method for evaluating the umami intensity of a monosodium glutamate solution, comprising the following steps:

[0045] Step 1: Take a certain amount of MSG (monosodium glutamate) granules and dissolve them in a solvent to obtain a sample MSG solution.

[0046] Step 2: Give the sample umami solution to the subject and mark it with time, and collect the EEG signal through multi-channel EEG.

[0047] Step 3: Preprocess the EEG signal, including downsampling, filtering, removing bad signals, re-referencing, and calculating time domain features and frequency domain features.

[0048] Step 4: Use the EEG time domain features and frequency domain features of the sample to be tested as input features, use the random forest algorithm model established by the standard concentration gradient sample solution to make predictions, and output the predicted value of the umami intensity of the sample to be tested.

[0049] Furthermore, the solvent in step 1 is distilled water, and the concentrations of the MSG solutions are 0.04 g / 100 ml, 0.16 g / 100 ml, 0.4 g / 100 ml, and 1.2 g / 100 ml, respectively. The downsampling in step 3 is from 500 Hz to 250 Hz, the filtering frequency range is 0.5-40 Hz, and the data capture period is 10 seconds after the marking in step 2. The frequencies are Delta band: 0.5-4 Hz, Theta band: 4-8 Hz, Alpha band: 8-13 Hz, and Beta band: 13-30 Hz.

[0050] The specific steps include:

[0051] 1) Preparation of MSG solution with standard concentration gradient

[0052] Sample sodium glutamate solution (0.04g / 100ml): Accurately weigh 0.04g of sodium glutamate into a 100ml sample bottle and dissolve it to the mark with distilled water to obtain a sodium glutamate solution with a concentration of 0.04g / 100ml.

[0053] Sample sodium glutamate solution (0.16 g / 100 ml): Accurately weigh 0.16 g of sodium glutamate into a 100 ml sample bottle and dissolve it to the mark with distilled water to obtain a sodium glutamate solution with a concentration of 0.16 g / 100 ml.

[0054] Sample sodium glutamate solution (0.4 g / 100 ml): Accurately weigh 0.4 g of sodium glutamate into a 100 ml sample bottle and dissolve it to the mark with distilled water to obtain a sodium glutamate solution with a concentration of 0.4 g / 100 ml.

[0055] Sample sodium glutamate solution (1.2 g / 100 ml): Accurately weigh 1.2 g of sodium glutamate into a 100 ml sample bottle and dissolve it to the mark with distilled water to obtain a sodium glutamate solution with a concentration of 1.2 g / 100 ml.

[0056] Preparation of pure aqueous solution

[0057] Sample aqueous solution: Add distilled water to the 100ml sample bottle to the scale to obtain distilled aqueous solution.

[0058] 2) Preparation before testing

[0059] Participants were required to complete a health questionnaire confirming their physical health, lack of mental, neurological, or psychological disorders, and a normal sense of taste. A pre-test health check was conducted to ensure compliance with test requirements. A sample was randomly selected from eligible participants to ensure representative results. To minimize experimental bias, participants with extensive taste testing experience were avoided, and participants were selected from diverse backgrounds, including age, gender, and occupation, to obtain comprehensive data on umami intensity.

[0060] Subjects are prohibited from taking foods and medicines that may affect their taste, such as antibiotics, vitamin supplements, etc., within 2 hours before participating in the test to ensure data accuracy and experimental safety.

[0061] 3) Instrument analysis conditions

[0062] Laboratory: The general testing environment temperature should be 20-25°C, with relative humidity controlled between 40% and 70%. The testing environment should be quiet and comfortable. Any factors that may interfere with EEG acquisition should be minimized.

[0063] Electrode requirements: Electrode placement should comply with the 10-20 international standard. The reference electrode CZ is located at the center of the top of the head.

[0064] Data collection: Assist the subject to wear the EEG cap and use conductive paste to reduce the electrode impedance to below 5 kΩ. Set the sampling rate to 500 Hz.

[0065] 4) Testing process

[0066] Using distilled water as the solvent, MSG solutions of varying concentrations were prepared as taste stimulus samples. All samples were maintained at a constant temperature of 21°C to eliminate the effects of temperature differences on the results. During the test, sample solutions were provided to the subjects in a pseudo-random order. Before each tasting, the subjects calmed down and, after 3 ml of the solution entered their mouths at a constant speed, savored the taste carefully for 15 seconds, keeping their eyes closed and their heads and bodies still. After the tasting, the subjects spit out the sample according to the prompts and rate the umami intensity of the solution on a 9-point scale. The subjects then rinsed their mouths repeatedly with clean water and rested for at least 3 minutes before tasting the next sample. Throughout the test, continuous EEG signals were recorded simultaneously, with at least four samples of each concentration collected in order to obtain a sufficient sample size to train the machine learning model.

[0067] 5) Data processing and model building

[0068] The EEG data was downsampled to 250Hz, and the filtering range was set to 0.5-40Hz. The EEG data for 10 seconds after the time stamp was selected and cut into 2-second segments as independent sample data to enhance the sample data. This method can increase the number of model training samples and improve model performance compared to direct feature extraction. Feature extraction was performed on each channel separately, and multiple different features were extracted, including time domain features: maximum signal amplitude (Maximum Value), minimum signal amplitude (Minimum Value), mean signal amplitude (Mean Value), median signal amplitude (Median), peak-to-peak value (Peak-to-Peak Value), average rectified value (Average Rectified Value), variance (Variance), standard deviation (Standard Deviation), kurtosis (Kurtosis), skewness (Skewness), and RMS amplitude (RMS Amplitude). Frequency domain features: Frequency Centroid, RMS Frequency, Frequency Standard Deviation, Spectral Skewness Mean, and Power Spectral Entropy. This approach provides a more comprehensive EEG feature set while increasing the model's complexity and parameters. A random forest algorithm was used to train the model, using all features from each channel as feature parameters to establish a classification model for umami intensity.

[0069] For classification tasks, random forests use majority voting:

[0070]

[0071] Where: 1(·) is the indicator function, when h m (x) takes the value 1 when it is c, otherwise it is 0; c is the total number of possible categories;

[0072] First, we use a feature selection method based on mutual information (MI), combining SelectKBest and mutual_info_classif in Python, to automatically select the features most relevant to the target variable and remove redundant and irrelevant features, thereby improving model performance and efficiency. This method improves the quality of input features and also provides a better feature subset for subsequent random forest training, reducing subsequent computational complexity. Mutual information measures the mutual dependence between two random variables and is mathematically defined as:

[0073]

[0074] Where P(x,y) represents the joint probability distribution of X and Y, and P(x) and P(y) are the marginal probability distributions of X and Y, respectively. Mutual information I(X;Y) reflects how much information X provides to reduce the uncertainty of Y. A larger value indicates a closer relationship between X and Y.

[0075] Assume that the original dataset has N features:

[0076] X={X1,X2,…,X N}

[0077] After calculating the mutual information of all features, they are summed up by the mutual information value I(X i ; Y) Sort in descending order and select the first k = 30 features:

[0078]

[0079] in:

[0080]

[0081] On this basis, the variance-based adaptive tree depth control is combined in the random forest training process, so that the model can dynamically adjust the complexity to avoid overfitting, while improving the classification ability of umami intensity. The highly correlated features screened out by MI mutual information complement the deep optimization strategy of random forest, so that the model can not only efficiently learn effective information, but also improve the overall classification performance by adaptively adjusting the complexity. In the hyperparameter optimization stage, GridSearchCV is used for multi-model hyperparameter optimization to automatically select the optimal hyperparameter combination. During the training process, the decision tree structure is dynamically adjusted to optimize the number of features (such as max_features) and depth (such as max_depth) of the tree. For the decision tree model f(X;θ), our goal is to find the optimal hyperparameter θ * , so that the performance indicators (such as accuracy) under cross-validation (CV) are maximized:

[0082]

[0083] Where θ is a set of hyperparameters (such as max_depth, max_features), K is k-fold cross validation (usually K = 5), X k ,Y k is the k-th fold training data and label, Score is the model performance evaluation indicator (such as accuracy, F1 score, etc.), θ * is the optimal hyperparameter after optimization. In this process, feature selection and hyperparameter optimization work together to enable the model to perform better classification results in complex tasks.

[0084] In this paper, the label refers to the actual umami solution intensity score, which uses a score range of 1 to 9. GridSearchCV is responsible for comprehensively searching the model's hyperparameters, including hyperparameters such as the maximum depth of the decision tree, and selecting the optimal hyperparameter combination. Meanwhile, the variance-based depth adjustment strategy automatically determines whether to continue splitting by calculating the variance of each node, thereby dynamically adjusting the actual depth of the decision tree.

[0085] The depth of the decision tree is generally fixed, and the training process depends on the Gini index:

[0086]

[0087] Too deep a tree may lead to overfitting, while too shallow a tree may reduce the expressiveness of the model. Depth is generally controlled by manually set hyperparameters. Adaptive depth control is performed using variance-based depth adjustment. Therefore, variance-based depth adjustment defines an adaptive depth control:

[0088] where yi Is the true label is the predicted value D t Is the optimal depth of the current decision tree when the error D t When it is less than a certain threshold e, it stops growing: if D t <ε. This method can automatically control the depth of the tree and avoid overfitting. The highly relevant features screened by MI are fully utilized in the optimized decision tree structure, further improving the classification effect of the overall model.

[0089] In this paper, variance refers to the variance of sample features within each decision tree node, reflecting the degree of dispersion of the sample data at that node. Specifically, when the data distribution is relatively concentrated, indicating that the sample features of the nodes have little variance, splitting can be stopped and the depth of the decision tree can be limited. When the data distribution is relatively dispersed, splitting can continue and the depth of the tree can be increased.

[0090] 6) Data Prediction

[0091] The newly collected data are input into the established model for prediction to obtain the umami intensity classification results of the corresponding umami solution.

[0092] The present invention is further described in detail below with reference to implementation cases.

[0093] Example 1

[0094] Instruments: 32-channel electroencephalogram (Neuracle NeuroHUB, China); analytical balance (sensitivity 0.0001 g, G&GJJ1000).

[0095] Analysis Conditions: A total of 20 participants participated in the test, including 10 males and 10 females, aged 18-35 years. All participants were in good physical condition and had passed a taste screening. The MSG solution concentrations were 0.04g / 100ml, 0.16g / 100ml, 0.4g / 100ml, and 1.2g / 100ml. EEG data were collected using a 32-lead EEG system with a 10-20 international standard lead system and a sampling rate of 500Hz. Before each tasting, participants calmed down and, after a 3ml sample was introduced into their mouths at a constant rate, savored the sip for 15 seconds, keeping their eyes closed and their head and body still. After tasting, participants were prompted to spit out the sample and rate the umami intensity of the solution on a 9-point scale. They then rinsed their mouths repeatedly with water and rested for at least 3 minutes before tasting the next sample. Continuous EEG signals were recorded throughout the test, with at least four samples collected for each concentration.

[0096] The data processing process is as follows:

[0097] The data was downsampled from 500Hz to 250Hz and band-pass filtered from 0.5 to 40Hz. The EEG data 10 seconds after the time stamp was selected and cut into 2-second segments for data enhancement. Feature extraction was performed on each channel. A total of 1515 training samples were obtained, each with a dimension of 640. The trained random forest model was optimized to obtain the umami intensity classification of different umami solutions.

[0098] First, a feature selection method based on mutual information (MI) is used to select the first 30 features as training samples. The final selected features are: FC2_F1_Maximum_Value, FC2_F3_Mean_Value, FC2_F4_Median, FC5_F3_Mean_Value, FP1_F1_Maximum_Value, FP1_F2_Minimum_Value, FP1_F3_Mean_Value, FP1_F4_Median, FP1_F6_Average_Rectified_Value, FP1_F11_Root_Mean_Square, FP1_F13_RMS_Amplitude, C3_F1_Maximum_Value, C3_F2_Minimum_Value e. C3_F3_Mean_Value, C3_F4_Median, C4_F1_Maximum_Value, C4_F2_Minimum_Value, C4_F3_Mean_Value, C4_F4_Median, CP2_F3_Mean_Value, CP5_F1_Maximum_Value, CP5_F2_Mini mum_Value, CP5_F3_Mean_Value, CP5_F4_Median, FP2_F1_Maximum_Value, FP2_F2_Minimum_Value, FP2_F3_Mean_Value, FP2_F4_Median, P3_F3_Mean_Value, P4_F1_Maximum_Value.

[0099] Among the features selected above, the feature naming follows the combination rules of electrode position and frequency band features, where the prefix indicates the electrode position (such as FC2, FP1, C3, etc.), the middle letter indicates the frequency band (such as F1, F2, F3, etc.), and the suffix indicates the statistic type of the feature (such as Maximum_Value, Mean_Value, Median, etc.), so as to clearly distinguish different electrodes, frequency bands and statistical features.

[0100] In this embodiment, the 30 features selected are derived from the time domain features and frequency domain features of each channel in the same EEG sample. These features are derived from the time domain and frequency domain data of the same sample, ensuring that the selected features can reflect the signal feature relationship within the sample.

[0101] By adjusting the depth based on variance, we calculated the variance of the cross-validation scores at each depth and obtained the optimal depth (max_depth = 50). At the same time, we used GridSearchCV to optimize the model hyperparameters and automatically selected the optimal hyperparameter combination: max_features = 'sqrt', min_samples_leaf = 2, min_samples_split = 4, and n_estimators = 400.

[0102] The EEG data of umami solutions with unknown umami intensity were input into the trained model to predict umami intensity, with a prediction accuracy of 97.03%.

[0103] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0104] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for evaluating the umami flavor intensity of MSG based on EEG and machine learning, characterized in that: The following steps are involved: Obtain brain wave data after testers taste a sample of a specific concentration of MSG solution; Preprocess the EEG data and extract EEG features, including multiple time domain features and frequency domain features; A feature selection method based on mutual information is used to select multiple time domain features and frequency domain features of brain waves to obtain a feature data set; Dividing the feature dataset into a training set and a test set; The optimized random forest model is used to train the training set and output the corresponding MSG umami intensity; The trained random forest model is used to predict the umami intensity of the MSG solution sample to be tested.

2. The method for evaluating the umami flavor intensity of MSG based on EEG and machine learning according to claim 1, characterized in that: Preprocessing of brain wave data includes the following steps: Downsample the EEG data to 250Hz; Filter the 250Hz EEG data, and set the filter range to 0.5-40Hz; The filtered EEG data were time-stamped and segmented at 2-second intervals, with each segment of EEG data being treated as an independent sample.

3. The method for evaluating the umami flavor intensity of MSG based on EEG and machine learning according to claim 1, wherein: Extract EEG features, including: Extract the time domain features of the preprocessed EEG data, including the maximum signal amplitude, minimum signal amplitude, average signal amplitude, median signal amplitude, peak-to-peak value, signal variance, signal standard deviation, signal kurtosis, signal skewness, RMS amplitude, and average rectified signal value; The preprocessed EEG data is converted into a frequency domain signal using Fourier transform, and multiple frequency domain features of the frequency domain signal are extracted, including frequency centroid, RMS frequency, frequency standard deviation, spectral skewness average, and power spectrum entropy.

4. The method for evaluating the umami flavor intensity of MSG based on EEG and machine learning according to claim 1, wherein: The feature selection method based on mutual information is used to select multiple time domain features and frequency domain features of brain waves, which specifically includes the following steps: Calculate the mutual information of all time domain features and frequency domain features of the same EEG data sample; All the mutual information values obtained are sorted in descending order, and the first 30 features are selected as the selected features.

5. The method for evaluating the umami flavor intensity of MSG based on EEG and machine learning according to claim 4, characterized in that: Calculate the mutual information of all time domain features and frequency domain features of the same EEG data sample, including the following formula: Among them, P(x,y) represents the joint probability distribution of feature X and feature Y, and P(x) and P(y) are the marginal probability distributions of feature X and feature Y respectively.

6. The method for evaluating the umami flavor intensity of MSG based on EEG and machine learning according to claim 1, wherein: During the training of the random forest model, GridSearchCV is used to optimize the model hyperparameters, specifically including the following formula: Among them, θ is a set of hyperparameters, including the number of features max_features and depth max_depth of the random forest model; K represents the total number of k-fold cross validation; X k ,T k Represents the k-th fold training data and true label value respectively; Score is the performance evaluation index of the random forest model, θ * is the optimal hyperparameter after optimization.

7. The method for evaluating the umami flavor intensity of MSG based on EEG and machine learning according to claim 1, wherein: During the training process of the random forest model, a variance-based depth adjustment strategy is adopted to dynamically adjust the depth of the random forest model decision tree.

8. The method for evaluating the umami flavor intensity of MSG based on EEG and machine learning according to claim 7, characterized in that: The variance-based depth adjustment strategy is used to dynamically adjust the depth of the random forest model decision tree, including: Calculate the variance of sample features in each decision tree node of the random forest model; When the variance of the sample features in a decision tree node is less than the set stop threshold, the decision tree node stops splitting; When the variance of the sample features in a decision tree node is greater than the set split threshold, the decision tree node continues to split.

9. A system for evaluating the umami flavor intensity of MSG based on EEG and machine learning, comprising a computer system, characterized in that: When executed, the computer system can implement the MSG umami intensity evaluation method based on EEG and machine learning as described in any one of claims 1 to 8.

Images